AI Implementation Services ROI: What Works, What Doesn’t

Coy Cardwell
Principal Engineer

The first thing to understand when talking about AI in the context of solution development is that you can use AI as one of the many tools in the development of the solution, and/or use AI within the solution itself. These are important distinctions and they are not mutually exclusive.

Saying we provided a solution using AI is a bit like saying we built something with a hammer. Accurate, but it really does not convey anything meaningful besides acknowledging we used tools, it is no different than saying ‘we used code’ to build software.

Saying we built a solution that leverages AI is a bit like saying we made a ‘smart’ product. Again, while correct, there is very little substance to the message. Simply saying that we used AI within a solution is like saying we leveraged an external API for specific workflow tasks.

We know the difference.

As is common with new and complex technologies, companies do not have the resources to deeply learn each new technology and must rely on specialists. For us, generative and agentic AI tools and technologies are our business as we are solution developers, we have the resources to adopt these new technologies so we can effectively use them for our clients.

Many companies find that having a technology partner to design and/or implement the integration paths of these new technologies is the best approach, so they can focus on what makes their business succeed and we can help them implement the right technical tools and solutions.

Here I want to talk about research and data from Gartner, McKinsey, and BCG around what AI implementation actually involves, the pitfalls along the path, and how to decide the age-old question of build, buy, or reuse. Does your company have the ability to leverage AI properly internally with current tools? Should you just buy an AI tool? Do you need an external vendor that can implement with better ROI?

What Does an AI Implementation Service Do?

AI in all its forms is just a tool. A powerful tool, but still just a tool. Like any tool, what matters is the person using the tool! Do they have the right technical experience and outlook for tasks at hand? Do they understand what they are building from a business perspective?

Based on our real world experiences with clients, these are the right questions to ask yourself as you build using agentic tools, as well as when implementing agentic solutions.

This is where real implementation work happens, in the space between understanding the business problem and articulating that properly so the problem can be solved with technology. In this instance an AI implementation partner provides the specialized skills and experience required to bridge the gap from idea to solution, instead of another ‘failed corporate initiative’.

Leveraging AI requires changes in how components and technologies are deployed and integrated, but it does not change the reality that all the activities around building production level solutions must still be addressed, and that new procedures must be implemented as well.

Agentic solutions are dynamic and built on your ever-changing business data. That means building and maintaining current, accurate data pipelines that feed the solutions workflow, without having to rebuild or entirely reengineer the data; intelligently integrating AI tools and agentic solutions into existing workflows, databases, and applications; enhancing automation; enabling proper governance controls and audit logging; monitoring for when things start drifting; training internal teams to evaluate what the AI produces rather than just accepting it.

This is the kind of work we do, in the weeds, making sure technology works for you.

What these services don’t do is replace business acumen. We help organisations figure out where AI actually helps, through reducing operational burden and increasing ROI vs. where it adds cost or complexity without benefit. Together we find a balance between these ideas.

Why Do Most AI Projects Fail to Deliver ROI?

The failure rate isn’t an edge case. It’s the baseline.

Only 28% of enterprise AI use cases fully meet ROI expectations. Gartner, April 2026 The primary success factor was workflow integration not things like model selection.

BCG’s 2025 Build for the Future study surveyed 1,250 global enterprises. Only 5% qualify as genuine AI value leaders. The other 60% report minimal gains despite significant investment. BCG calls this the “widening AI value gap”, a growing divide between organisations that have embedded AI into core operations and those permanently stuck in pilot mode bringing experiments into production.

60% of enterprises report no material value from AI. Only 5% achieve value at scale. BCG, Sept 2025 The 5% are 1.7x more likely to see revenue growth and 1.6x higher EBIT margins.

McKinsey’s 2025 State of AI survey found that 88% of organisations now use AI in at least one function. Only 39% report any measurable EBIT impact. Two-thirds are stuck experimenting, piloting, never actually scaling their efforts.

The consistent finding across all three of these sources is that adoption itself isn’t the problem, but it’s the implementation, the ability to move from a working prototype to a working system. Companies do not have the internal capability to take on this new technology directly, without strategic partners, and struggle with that reality as they have from previous advancements. Forrester’s AI predictions for 2025 found that three out of four firms attempting to build advanced AI architectures independently will fail. This is not from lack of trying, but because the evolving details around model integration, RAG stacks, and enterprise-grade reliability require specialised experience that most engineering teams simply haven’t yet accumulated.

The Honest Business Case for AI in 2026

“Saying we solved the problem with AI is like saying we used a hammer. It really is just another tool — a more advanced one, but still just a tool.”

Coy Cardwell
AI Principal Engineer, First Line Software

Here’s my actual opinion, after more than a decade evaluating technology for business ROI. Generative AI solves one (albeit important) problem meaningfully that we all face. Time to market. That’s it.

Others see it differently, and that’s fine. But in practice, what AI does is save time, if you use it correctly in the context of the problem you are trying to solve. If not, like any tool misapplied, it will take longer and cost more than a more traditional alternative approach. Faster prototyping, faster drafting, faster orientation in field or territory in which your team lacks. Gains of this nature are real, but can be fragmented and hard to measure. Capturing and measuring these gains as they compound across a project lifecycle can be the difference in a successful addition to the business or just another failed experiment.

At least 30% of GenAI projects will be abandoned after proof of concept due to poor data quality, escalating costs, or unclear business value. Gartner, 2024 Early adopters reported average productivity gains of 22.6% when implementation was correctly scoped.

Used correctly, AI reduces time on well-defined work that can be quantified. Used incorrectly, it typically costs more than the alternative while providing an inferior result. The tools don’t know what “done” looks like and that’s where curating that definition is the job.

Who Needs AI Implementation Services? Two Stories We See.

The businesses that come to us tend to fall into one of two situations I see regularly.

The tech lead who got stuck

Someone, convinced that AI could build “the new thing” end-to-end and, spoiler alert, it couldn’t. This is typically not because the AI failed, but because while building complete solutions still requires understanding the full stack, the integrations, the things that break in production at 2 a.m., it also requires proficiency in this new technology. They’d gotten far. Then they hit something outside their wheelhouse, and the whole thing stalled. They call us and ask “can you do this?”

What they need isn’t more AI. They need end-to-end solution knowledge, a partner who could take AI-generated and/or agentic components and integrate them into the systems the tech lead already understood, that actually run the business, without breaking the workflow or the bank.

The one who didn’t know what to ask

Leadership could ask AI anything. They just didn’t know enough about their data to ask the right questions or evaluate the answers properly, so they couldn’t recognize when the output was noise. Like the early Google users who had access to all the world’s information but couldn’t formulate the right query to get the results they really needed. So the tool was generating things but nobody could evaluate whether any of it was actionable.

What they needed was experience in knowing where AI genuinely reduces the drudge work, and the discipline to keep the tools focused on business value. For most organisations, honestly, it’s still easier to pay people to do some things than to maintain a new technology. That’s not a criticism, it’s just the realistic calculation most businesses have to make, so applying AI correctly becomes the task.

McKinsey research confirms that organisations reporting significant financial returns from AI are twice as likely to have redesigned end-to-end workflows before selecting tools, not after. Tool selection comes last. You might build it with AI, you might use AI in the solution, you might do both, but to do so without a plan is not good business.

Build It Yourself or Bring Someone In?

Three things determine this: how well your team understands the existing system, how fixed the timeline is, and whether you have people who can evaluate AI outputs end-to-end. Not just generate them and evaluate them for accuracy as well as business value.

Factor	Build it yourself	Bring in outside help
When it works	Full-stack team that can evaluate AI output end-to-end	Fixed timeline, complex integration, or team hasn’t done this before
Time to value	Slower, learning curve baked in	Faster, we’ve already made the mistakes
Risk	High if there are system knowledge gaps	Lower as we find the failure points before they find you
Real cost	Low upfront but expensive if you have to rework it	Higher upfront but cheaper than an abandoned project
Best for	Greenfield work with strong internal engineering	Complex integrations, regulated environments, hard deadlines

The cost question is almost always framed incorrectly. The upfront number is visible. The cost of a project that stalls six months in and gets quietly abandoned rarely shows up anywhere except in a meeting where someone is let go, or a business leader is explaining where the budget went to shareholders in an uncomfortable meeting. That’s the number that actually matters.

Where AI Implementation Reduces Costs vs. Where It Doesn’t

ROI from AI comes from a combination of properly curating the type of work being automated and the evaluation of what is produced by the right business owner. If the work produced by the AI tools cannot be properly verified, then leveraging it at scale becomes a business risk.This is a business judgment problem, not a technical issue.

In my experience with clients, the pattern has been consistent. AI generates the most value and ROI when experienced engineers are involved from the beginning. Properly planning the work the AI tools and agents are being tasked to do by scoping the problem correctly and then dividing the work properly between AI and humans is the core of a great AI implementation team. The rest is knowing your own business well enough to work with the AI team to find the places where AI can do the right work, and having the discipline to know when to avoid AI.

What Separates AI Projects That Last from Those That Don’t?

Gartner’s 2025 research on AI maturity found that in high-maturity organisations, 45% of AI initiatives stay in production for three or more years. In low-maturity organisations: 20%. The differentiator isn’t the model, it seems to be the structure of the effort. Factors like dedicated AI leadership, centralised governance, and integration into real workflows.

High-maturity AI organisations are 2.25x more likely to keep AI projects in production for 3+ years. Gartner, 2025 Trust and workflow integration are the primary differentiators.

What we sell at First Line Software isn’t AI. It’s experience in implementing and maintaining complete systems that run businesses. That’s always been the job. AI is just the current new tool in a long line of new tools, and while everyone else wants to talk about today’s toy, I’m more interested in long-term ROI and what’s still working in three years.

BCG’s research recommends a “10-20-70 rule” for successful AI implementation: 70% of effort on people and process, 20% on technology, 10% on the algorithms. BCG, 2025. Most organisations have this completely inverted, thinking they can just hand the ball to AI.

“AI tools will change — there could be a better one tomorrow. What doesn’t change is the need for people who know how to build systems that actually work.”

Coy Cardwell
AI Principal Engineer, First Line Software

FAQ

What’s the difference between buying an AI tool and using AI implementation services?

Buying an AI tool gives you access to a tool just like buying a hammer. Implementation services provide the engineering experience and ability to connect AI tools to your data, your systems, and your workflow. The tool produces outputs. Implementation determines whether those outputs are accurate, integrated, running reliably, and providing business value, not just a demo. Most ROI failures happen in this gap between acquiring a tool and realizing ROI.

How long does a typical AI implementation project take?

In terms of business scope, integrating a workflow with data integration leveraging AI tools can take less than a week if your business can handle that speed, we can work in stints as short or as long as required, from three days to months. We call it RACE.

When we talk about implementing AI tools within your business the time involved is more about the planning necessary to ensure the solution fits the business requirements. Working with you to understand the business and knowing where and how to integrate AI and making data accessible for real, targeted decisions realistically takes time for humans to review and approve. While we can use our RACE methodology on the development and deployment, it typically takes 4 to 12 weeks depending on the extent of the data to be wrangled, its readiness, and existing system complexity, due to approval and acceptance testing. Extensive, end-to-end implementations covering pipeline build, monitoring, and compliance controls can run 3 to 6 months in enterprise environments. Longer timelines usually signal a data or process issue, not an AI tool or speed of development problem.

Why do so many AI proof-of-concept projects fail to reach production?

Lack of knowledge and experience in end-to-end SDLC practices.

Edge cases and integration complexity nobody thought about lead the charge here. Gartner estimates at least 30% of GenAI projects are abandoned after proof of concept. That’s not an AI tool failure, it’s a systems integration failure, and it’s preventable.

Is it better to build internally or use an external AI implementation partner?

Companies deploy small scale projects all the time within their organizations, I used to be one of the people doing that. However, I worked for software companies that had the knowledge and experience to get projects from zero to hero, or my IT experiments would never have seen the light of day. If your IT department supports the business well, it does not mean they are in the business of redesigning software systems, typically they are great at maintaining what you have over building things from scratch. They have the right ideas, but need the right help implementing them.

What does a successful AI implementation actually look like?

It provides business value and ROI while it stays in production. Value can be realized in different forms, from measurably reducing a specific type of work, seamlessly enhancing an existing workflow, or monitoring for output quality, to self correcting as conditions change. Finished and effective, it turns out, is the hard part, because no one defines what success is early.

How do you measure ROI from AI implementation services?

Do the new tools save you more time than they cost? That’s the ROI.

Automated tasks are the easiest to measure for most companies; did they save time and was that time more effectively spent on human tasks? The number of hours per week spent on document drafting is a good example. We know the tools can speed code production, which means faster time to market. Error rates in a specific process can be quantified and addressed. Broad productivity claims are difficult to verify. Metrics for measuring success should be agreed before implementation begins, or you will end up trying to justify the cost of a failed project.

Key Terms

Term	What it actually means
AI implementation services	The engineering work that moves an idea leveraging AI from a dream into something that actually enhances business. Data pipelines, system integration, governance, maintenance. AI tool choice is the easy part.
Proof of concept (POC)	A limited trial in a controlled environment. Typically not fully reflective of real operating condition variables such as data, edge cases, and new compliance requirements. The gap between POC and production is where most projects die based on current data.
Workflow integration	Leveraging AI outputs into existing business processes and systems so they’re used directly in operations, when a human is not required to accomplish the given task.
RAG (Retrieval-Augmented Generation)	An architecture where the AI process retrieves relevant data at inference time rather than relying only on what it learned in training. Required in enterprise contexts where the answer depends on your data. Done wrong and you are making a decision on bad info.
AI hallucination	Plausible-sounding output that is factually wrong. The primary risk in any deployment where a human isn’t reviewing the output or where the system isn’t designed to catch it.
Systems integrator	An engineering firm or team that connects software, data sources, and applications into a complete functioning system. In the context of AI, the teams who make the AI tools actually work for your business, inside your environment, not just in a PoC.
AI maturity	How deeply an organisation has embedded AI into operations, governance, and decision-making in a way that holds up over time. High-maturity organisations keep AI in production more than twice as long as low-maturity ones (Gartner, 2025).

Is AI Implementation Right for Your Operation?

We don’t start with a demo or technology. We start with your business. What you are doing, what you’re trying to increase, change, or improve. Then we talk about how AI implementation services would actually close the gaps you have or if the problem to be solved just needs to be done quickly, we can help you use AI in the right way for your business.

That conversation is usually more useful than a technology overview.

Discuss your current systems and where AI fits →